Failure pattern
Missing isolation
Environments, tenants or data that should have been separate were reachable from one another.
Preventive controls from the record
| Incident | Control |
|---|---|
| INC-2026-0007 WorkOS (affected organization) | Run offensive agents behind enforced egress allowlists, and keep internal tools off the internet with strong credentials, rate limits and volume alerts. |
| INC-2026-0004 Anthropic | Verify every network path out of an evaluation environment before a run, enforce egress denial at the network layer, and monitor runs for public IP addresses in real time. |
| INC-2026-0003 PocketOS | Scope infrastructure tokens to one environment and operation, keep them out of the repository, and store backups apart from the data. |
| INC-2025-0008 Replit | Separate production credentials and databases from development environments, and require approval for destructive commands. |
| INC-2025-0005 Asana | Apply the same tenant-isolation tests to MCP and API surfaces as to the core product before launch. |
| INC-2024-0005 OpenAI | Roll infrastructure changes out in phases with control-plane health checks, and keep an access path that works when the control plane does not. |
| INC-2023-0001 OpenAI | Check at the point of use that cached data belongs to the requesting user, and roll server changes out in stages with error-rate monitoring. |
Controls are editorial summaries by Incidents.si, not statements by the organizations involved.
7 incidents with this pattern
- A researcher’s AI hacking agent leaves its test environment and breaches a WorkOS tool
WorkOS found unauthorized bulk access to an internal tool used by its sales team. A security researcher later said their ARTEX agent had escaped its test environment during benchmark runs and attacked WorkOS unintentionally.
WorkOS (affected organization) · Bulk retrieval of business contact details, account information and excerpts of customer communications from an internal go-to-market tool
- Claude models break into real company systems during misconfigured cyber evaluations
Anthropic disclosed that Claude models, told they were in an offline simulation, reached the real internet during cyber evaluations and compromised the infrastructure of several organizations, including through a malicious PyPI package.
Anthropic · Unauthorized access to real third-party systems in four incidents; a malicious PyPI package ran on 15 real systems
- Coding agent deletes PocketOS production data and its backups in seconds
While working in staging, a coding agent used an account-wide Railway token to delete a volume. It was the production volume, and the backups were stored inside it.
PocketOS · Production database volume and its backups deleted; newest usable backup was three months old
- Replit’s coding agent deletes a production database during a code freeze
During a user’s test of Replit’s coding agent, the agent deleted the production database despite a code freeze. Replit’s chief executive called it unacceptable and announced safeguards.
Replit · Production records for over 1,200 executives and nearly 1,200 companies deleted; later recovered
- Asana takes its MCP server offline after a cross-organization data exposure bug
A bug in Asana’s MCP server could have exposed information from one customer’s workspace to MCP users at other customers. Asana disabled the server for nearly two weeks.
Asana · Data from one organization potentially visible to MCP users in other organizations; server offline from June 5 to June 17
- New telemetry service takes down OpenAI’s API, ChatGPT and Sora for hours
A new telemetry service caused every node in OpenAI’s clusters to make heavy Kubernetes API calls, which took down the control plane and, with it, all OpenAI services.
OpenAI · All OpenAI services degraded or unavailable from 3:16 PM to 7:38 PM PST
- ChatGPT bug exposes chat titles and payment details of other users
OpenAI took ChatGPT offline after a caching bug let some users see titles from other users’ chat histories and, for some subscribers, payment-related details.
OpenAI · Chat titles shown to other users; payment details of 1.2% of active ChatGPT Plus subscribers potentially visible during a nine-hour window
No incidents match these filters.